Estimation device, estimation method, and trained model generation method

Abstract

In the present invention, an image input unit receives input of image data showing a gel structure with a curved portion. An estimation unit estimates a parameter related to the gel structure on the basis of the input image data by using a trained model that is trained to receive input of image data showing a gel structure to output a parameter related to the gel structure shown in the image data.

Description

Estimation device, estimation method, and trained model generation method

[0001] The present invention relates to an estimation device, an estimation method, and a method for generating a trained model.

[0002] In recent years, medical and healthcare applications of materials with bio-like properties, such as hydrogels, have been attracting attention (Non-Patent Document 1). Gel structures made from gels such as hydrogels are expected to be used in a wide range of fields in the future, and there is a demand for controlling the shape of gel structures in various fields. As a method for controlling the shape of a gel structure, Patent Document 1 discloses a method for controlling the shape of a gel structure by buckling delamination of a laminate.

[0003] Japanese Patent Application Laid-Open No. 2020-062843

[0004] Riku Takahashi, Aya Tanaka, Masumi Yamaguchi, Yukio Koike, "Bionics Technology Opening Up the Future of Medical Health," NTT Technical Journal, Nippon Telegraph and Telephone Corporation, May 2021, pp. 32-36. Alhassan Mumuni, Fuseini Mumuni, "Data augmentation: A comprehensive survey of modern approaches," Array, vol. 16, 2022.

[0005] Quantifying the characteristics of a gel structure made of a material such as hydrogel requires manual measurement. It is also difficult to determine the parameters specified at the time of design from the manufactured structure. An object of the present invention is to provide an estimation device, an estimation method, and a trained model generation method that can easily estimate parameters related to a gel structure.

[0006] A first aspect of the present invention is an estimation device comprising an image input unit that accepts input of image data depicting a gel structure having a curved portion, and an estimation unit that estimates parameters relating to the gel structure based on the input image data depicting the gel structure using a trained model that has been trained to output parameters relating to the gel structure depicted in the image data depicting the gel structure.

[0007] A second aspect of the present invention is an estimation method comprising the steps of: accepting input of image data depicting a gel structure having a curved portion; and estimating parameters relating to the gel structure based on the input image data depicting a gel structure having a curved portion using a trained model trained to output parameters relating to the gel structure depicted in the image data depicting the gel structure having a curved portion.

[0008] A third aspect of the present invention is a method for generating a trained model, comprising the steps of: acquiring a training dataset including a combination of sample image data depicting a gel structure having a curved portion and parameters related to the gel structure depicted in the sample image data; using the training dataset to learn internal parameters of a machine learning model so as to output parameters related to the gel structure from image data depicting the gel structure; and generating a trained model using the trained internal parameters, wherein the trained model causes a computer to execute the steps of accepting input of image data and calculating parameters related to the gel structure depicted in the image data based on the input image data and the internal parameters.

[0009] According to the above aspect, the parameters related to the gel structure can be easily estimated.

[0010] FIG. 1 is a diagram showing an example of a laminate to be estimated according to an embodiment; FIG. 2 is a schematic diagram showing the configuration of an estimation device according to a first embodiment; FIG. 3 is a diagram showing an example of image data extension according to the first embodiment; FIG. 4 is a flowchart showing a method for training a machine learning model by the estimation device according to the first embodiment; FIG. 5 is a flowchart showing a method for estimating parameters by the estimation device according to the first embodiment; FIG. 6 is a confusion matrix showing experimental results according to the first embodiment; FIG. 7 is a graph showing experimental results according to the second embodiment; FIG. 8 is a schematic diagram showing the configuration of an estimation device according to a third embodiment; FIG. 9 is a diagram showing an example of presentation of evidence according to the third embodiment; FIG. 10 is a schematic diagram showing the configuration of an estimation device according to a fourth embodiment; and FIG. 11 is a schematic block diagram showing the configuration of a computer according to at least one embodiment.

[0011] <About Estimation Target> In the following embodiment, parameters related to a laminate L having a gel structure L2 with a predetermined three-dimensional structure are estimated. Fig. 1 is a diagram showing an example of the laminate L to be estimated according to the embodiment. The laminate L to be estimated includes a solid substrate L1 and a gel structure L2.

[0012] <<Solid Substrate L1>> The modulus of rigidity of the solid substrate L1 is higher than that of the gel structure L2. Examples of organic materials forming the solid substrate L1 include polymeric materials and elastomers. Examples of polymeric materials include thermoplastic resins such as polyvinyl chloride, polystyrene, ABS resin, and polylactic acid, and thermosetting resins such as polyimide and phenolic resin. Examples of elastomers include polysilicone and synthetic rubber. The solid substrate L1 formed from an elastomer is easily deformed in response to stress. In a laminate L including such a solid substrate L1, the gel structure L2 can be deformed in response to deformation of the solid substrate L1. The solid substrate L1 can also be formed from a hydrogel with a swelling degree different from that of the hydrogel forming the gel structure L2 described below. Various additives may be added to the organic materials described above to add various functions to the solid substrate L1 based on the physical properties of the additives.

[0013] Examples of inorganic materials that are used to form the solid substrate L1 include glass, which has excellent transparency and chemical stability, conductors that generate heat when electricity is passed through them, magnetic metal bodies that generate heat when stimulated by a magnetic field, piezoelectric elements that generate electricity through stress, and light-emitting elements that emit light when electricity is passed through them.

[0014] (Gel Structure L2) The gel structure L2 is provided on one surface of the solid substrate L1.

[0015] Examples of polymeric materials constituting the gel structure L2 include water-soluble polymers such as polyacrylamide and polyvinyl alcohol, polysaccharides such as chitosan and alginic acid, and proteins such as collagen and albumin. These materials have a three-dimensional network structure and swell by absorbing a solvent to a large extent of their volume. A typical solvent that swells the polymeric materials constituting the gel structure L2 is water. In other words, the gel structure L2 may be a hydrogel.

[0016] The gel structure L2 may be composed of a polymer gel or a low molecular weight gel other than a hydrogel. Examples of other gels include organic polymer gels, inorganic gels, polysaccharide gels, protein gels, physical gels, chemical gels, lipogels, ion gels, microgels, and porous gels. The gel may also be a xerogel.

[0017] Various known methods can be used to synthesize the polymer material that constitutes the gel structure L2. For example, when the polymer material that constitutes the gel structure L2 is an acrylic polymer material, a three-dimensional network structure may be formed by crosslinking acrylic groups during polymerization of acrylic monomers.

[0018] The type of polymerization reaction when polymerizing the acrylic monomer is not particularly limited, but an example thereof is radical polymerization using a water-soluble photopolymerization initiator. Examples of the water-soluble photoinitiator include 2-oxoglutaric acid, 4'-(2-hydroxyethoxy)-2-hydroxy-2-methylpropiophenone (trade name: Irgacure 2959), lithium phenyl(2,4,6-trimethylbenzoyl)phosphinate (abbreviation: LAP), and 2,2'-azobis[2-methyl-N-(2-hydroxyethyl)propionamide] (trade name: VA-086).

[0019] During radical polymerization, a deoxidizer may be added to the reaction system to prevent polymerization inhibition by oxygen. Examples of the deoxidizer include a combination of glucose and glucose oxidase. Radical polymerization may also be carried out under an inert gas atmosphere such as nitrogen or argon.

[0020] When the polymer material constituting the gel structure L2 is a polysaccharide or a protein, the three-dimensional network structure may be formed by physical bonding of the polysaccharide or protein, or by crosslinking the polysaccharide or protein using a crosslinking agent, such as glutaraldehyde.

[0021] The strength of the gel structure L2 can be improved by increasing the crosslinking of the polymer material constituting the gel structure L2 by chemical crosslinking or physical crosslinking, or by increasing the concentration of the polymer material constituting the gel structure L2. For example, when preparing the gel structure L2 containing polyacrylamide by polymerizing an acrylamide monomer (precursor), the monomer concentration is preferably 0.8 mol / L to 8 mol / L, and more preferably 2 mol / L to 4 mol / L.

[0022] Furthermore, when methylenebisacrylamide is used as a chemical crosslinking agent when polymerizing acrylamide monomers, the concentration of the crosslinking agent is preferably 0.01 mol % to 2.0 mol % relative to the monomer, and more preferably 0.03 mol % to 1 mol %.

[0023] The gel structure L2 can contain various additives. The type of additive is not particularly limited as long as it does not inhibit the formation of the gel structure L2. Examples of additives include biomolecules that improve biocompatibility, silver nanoparticles or surfactants that provide antibacterial properties, ionic liquids or conductive polymers that increase conductivity, and magnetic nanoparticles that react to magnetic fields. Adding these additives to the gel structure L2 can impart any desired function to the hydrogel.

[0024] The upper surface of the solid substrate L1 is provided with an adhesive region L11 for fixing the gel structure L2 and a non-adhesive region L12 for not fixing the gel structure L2. For example, a silane coupling material having an adhesive functional group is applied to the adhesive region L11. The non-adhesive region L12 is sandwiched between the adhesive regions L11 and has an elongated shape.

[0025] The gel structure L2 has an adhesive portion L21 that is adhered to the solid substrate L1 and a non-adhesive portion L22 that is not adhered to the solid substrate L1. The adhesive portion L21 is a portion of the gel structure L2 that contacts the adhesive region L11 of the solid substrate L1 and is adhered to the solid substrate L1. The non-adhesive portion L22 is a portion of the gel structure L2 that contacts the non-adhesive region L12 of the solid substrate L1, and is not adhered to the solid substrate L1 because no silane coupling agent is present between the non-adhesive portion L22 and the solid substrate L1.

[0026] The volume of the gel structure L2 increases as it swells. At this time, the non-bonded portions L22 can freely increase in volume in the extension direction of the non-bonded portions L22 and in the direction away from the solid substrate L1 during the volume increase due to swelling. On the other hand, the non-bonded portions L22 are restricted from increasing in volume in a direction intersecting the extension direction of the non-bonded portions L22 because the bonded portions L21 are fixed to the solid substrate L1. As a result, the non-bonded portions L22 undergo buckling delamination, which causes them to bend significantly in the direction away from the solid substrate L1 to mitigate the increase in internal pressure caused by the volume increase. At this time, the non-bonded portions L22 are an example of a curved portion. As a result, a conduit is formed between the gel structure L2 and the solid substrate L1.

[0027] 2 is a schematic diagram showing the configuration of an estimation device 100 according to a first embodiment. The estimation device 100 according to the first embodiment estimates design parameters of a laminate L. Examples of the design parameters of the laminate L include the width of the non-bonded region L12 and the types and concentrations of chemical substances such as cross-linking agents that constitute the gel structure L2.

[0028] The estimation device 100 includes an image acquisition unit 101, a parameter input unit 102, a storage unit 103, a learning unit 104, an estimation unit 105, and an output unit 106.

[0029] The image acquisition unit 101 acquires image data showing the top surface (gel structure L2 side) of the laminate L. The image acquisition unit 101 may process the acquired image data to expand the learning dataset. FIG. 3 is a diagram showing an example of image data expansion according to the first embodiment. Examples of processing include rotation, inversion, and translation of the image data. The image acquisition unit 101 may also process the image data using the method described in Non-Patent Document 2. The parameter input unit 102 accepts input of design parameters of the laminate L shown in the image data acquired by the image acquisition unit 101.

[0030] The storage unit 103 stores a learning dataset D1 consisting of a combination of image data and design parameters, and a machine learning model D2. The machine learning model D2 has internal parameters and obtains the value of a response variable from the value of an explanatory variable by calculation using the internal parameters. The machine learning model D2 may be, for example, a deep learning model such as a CNN. The machine learning model D2 may also be a model implemented by transfer learning, in which the values ​​of the internal parameters are replaced with the output layer of an image classification model previously trained. Examples of pre-trained image classification models that can be used include Resnet50 and Efficient Net. The explanatory variables of the machine learning model D2 are image data, and the response variables are design parameters. The machine learning model D2 is implemented by a classification model. The output of the machine learning model D2 may be a label that defines each of multiple possible values ​​of the design parameters. For example, when estimating the width of the non-bonded portion L22, labels ranging from 0.1 μm to 1.0 μm in 2 μm increments may be output, or labels ranging from 1 μm increments may be output.

[0031] The learning unit 104 updates the internal parameters of the machine learning model D2 using the training dataset D1 recorded in the storage unit 103. For example, the learning unit 104 can update the internal parameters in the following procedure. The learning unit 104 inputs image data, which is an input sample of the training dataset D1, into the machine learning model D2 and performs a calculation using the internal parameters. The learning unit 104 calculates the error between the value obtained by the calculation and the design parameters, which are output samples of the training dataset D1. The learning unit 104 updates the values ​​of the internal parameters so as to reduce the error. When the machine learning model D2 is realized by a classification model, the design parameters, which are output samples, are represented as one-hot vectors.

[0032] The estimation unit 105 inputs the image data acquired by the image acquisition unit 101 into the trained machine learning model D2 and estimates the values ​​of the design parameters by performing calculations based on the internal parameters of the machine learning model D2. The output unit 106 outputs the values ​​of the design parameters estimated by the estimation unit 105 to a display device (not shown) or the like. The output unit 106 may also output the design parameters to a manufacturing device for the gel structure L2, causing the manufacturing device to manufacture a laminate L having characteristics similar to those of the laminate L depicted in the image data.

[0033] <Relationship Between Image Data and Design Parameters> The image data shows the gel structure L2 of the solid substrate L1. The non-bonded portions L22 of the gel structure L2 are curved due to buckling delamination. As shown in FIG. 3 , the non-bonded portions L22 meander in the extension direction. This is because a compressive load is applied to the non-bonded portions L22 in the extension direction due to an increase in the volume of the gel structure L2, causing buckling. Shape characteristics of the non-bonded portions L22, such as their height, width, and curvature in the thickness direction, as well as the amplitude and period of meandering, appear as shadows in the image data. The shape of the non-bonded portions L22 is determined by characteristic parameters of the gel structure L2, such as the hardness, thickness, degree of expansion, and width of the conduit. These characteristic parameters of the gel structure L2 are controlled by design parameters used when manufacturing the laminate L, i.e., the width of the non-bonded regions L12, and the types and concentrations of chemicals, such as cross-linking agents, that make up the gel structure L2. In other words, the hardness and water absorption of gel structure L2 are determined by the composition of the chemical substances that make up gel structure L2, and the height, width, and curvature of non-adhesive portion L22 are determined by the hardness and water absorption of gel structure L2 and the width of non-adhesive region L12, i.e., the distance between the fixed ends of non-adhesive portion L22. In this way, the shape characteristics of non-adhesive portion L22 that appear as shadows in image data, the characteristic parameters of gel structure L2, and the design parameters each have a correlation based on the causal relationship described above.

[0034] <Learning Method> FIG. 4 is a flowchart showing a learning method for a machine learning model by the estimation device 100 according to the first embodiment. First, the image acquisition unit 101 acquires multiple image data showing the top surface (the gel structure L2 side) of the laminate L and stores them in the storage unit 103 as input samples for the learning dataset D1 (step S1). The image acquisition unit 101 also processes the acquired multiple image data, performs an extension process, and stores them in the storage unit 103 as input samples for the learning dataset D1 (step S2). The parameter input unit 102 accepts input of design parameters for the laminate L shown in the multiple image data acquired by the image acquisition unit 101, and stores them in the storage unit 103 as output samples for the learning dataset D1 in association with the input samples (step S3). The parameter input unit 102 accepts input of, for example, the values ​​of the design parameters and the file names of image data showing the laminate L related to the design parameters. This allows the parameter input unit 102 to identify the image data corresponding to the input design parameters. The parameter input unit 102 associates the values ​​of corresponding design parameters with an input sample, which is image data having a specified file name, and with an input sample extended from the image data having the file name.

[0035] The learning unit 104 updates the internal parameters of the machine learning model D2 using the training dataset recorded in the storage unit 103 (step S4). For example, the learning unit 104 randomly divides the training dataset D1 into training data and validation data, updates the internal parameters using the training data, and verifies the accuracy using the validation data. The learning unit 104 may repeat the update and validation processes until the verified accuracy exceeds a certain level, or may perform the learning process for a predetermined number of epochs. The learning unit 104 records the trained machine learning model D2 in the storage unit 103 (step S5).

[0036] 5 is a flowchart showing a parameter estimation method performed by the estimation device 100 according to the first embodiment. First, the image acquisition unit 101 acquires image data showing the top surface of the laminate L (step S11). Next, the estimation unit 105 inputs the acquired image data to the trained machine learning model D2 stored in the storage unit 103 (step S12). The estimation unit 105 calculates values ​​of design parameters according to the internal parameters of the machine learning model D2 (step S13). In other words, the trained machine learning model D2 causes the estimation device 100 to execute a step of accepting input of image data and a step of calculating parameters based on the image data and the internal parameters.

[0037] The output unit 106 outputs the calculated values ​​of the design parameters (step S14).

[0038] Experimental Results The inventors actually estimated the design parameters of the gel structure L2 using the estimation device 100 according to the first embodiment. The machine learning model used in the experiment was a CNN classification model having, from the input side, a first convolutional layer, a first normalization layer, a second convolutional layer, a second normalization layer, a third convolutional layer, a third normalization layer, and a fully connected layer. The design parameter to be estimated was the width of the non-adhesive region L12. The labels associated with the output layer of the machine learning model were 200 μm, 400 μm, 600 μm, 800 μm, and 1000 μm. For training the machine learning model, 35 sets of training datasets were expanded by five sets per set through inversion and rotation, resulting in a total of 210 sets (6 × 35). FIG. 6 is a confusion matrix illustrating the experimental results according to the first embodiment. As shown in FIG. 6, the estimation device 100 according to the first embodiment can accurately estimate design parameters. The accuracy (number of correct answers / number of data) was 0.86, the precision (correct answer rate for each predicted value label) was 0.87, the recall (correct answer rate for each true value label) was 0.86, the F1 score (2*(Precision*Recall) / (Precision+Recall)) was 0.86, and the AUROC (Area Under the ROC) was 0.91.

[0039] <<Actions and Effects>> As described above, the estimation device 100 according to the first embodiment estimates design parameters of the gel structure L2 from the acquired image data using the trained machine learning model D2. As described above, there is a correlation between the image data and the design parameters of the gel structure L2. Furthermore, as shown by the experimental results, according to the first embodiment, the estimation device 100 can accurately calculate the design parameters of the gel structure L2 from the image data.

[0040] The machine learning model D2 according to the first embodiment is configured as a classification model, whereas the machine learning model D2 according to other embodiments may be configured as a regression model.

[0041] Second Embodiment An estimation device 100 according to a second embodiment estimates feature parameters representing features of a laminate L. Examples of the feature parameters of the laminate L include the height, width, and curvature of the laminate L in the thickness direction, as well as the amplitude and period of meandering.

[0042] The estimation device 100 according to the second embodiment has the same configuration as that of the first embodiment. The machine learning model D2 according to the second embodiment may be, for example, a deep learning model such as a CNN. The machine learning model D2 may also be a model implemented by transfer learning, in which the output layer of an image classification model in which internal parameter values ​​have been previously learned is replaced. As the previously trained image classification model, for example, Resnet50 or Efficient Net can be used. The explanatory variables of the machine learning model D2 are image data, and the objective variables are feature parameters. The machine learning model D2 is implemented by a regression model.

[0043] Experimental Results The inventors actually estimated feature parameters of the gel structure L2 using the estimation device 100 according to the second embodiment. The machine learning model used in the experiment was a CNN classification model having, from the input side, a first convolutional layer, a first normalization layer, a second convolutional layer, a second normalization layer, a third convolutional layer, a third normalization layer, and a fully connected layer. The feature parameter to be estimated was the amplitude of meandering in the non-adhesive region L12. For training the machine learning model, 35 training datasets were expanded by five sets per dataset through inversion and rotation, resulting in a total of 210 sets (6 × 35). FIG. 7 is a graph showing experimental results according to the second embodiment. As shown in FIG. 7, the estimation device 100 according to the second embodiment can accurately estimate feature parameters. The correlation coefficient of the experimental results shown in FIG. 7 was 0.62.

[0044] <<Actions and Effects>> In this way, the estimation device 100 according to the second embodiment estimates feature parameters of the gel structure L2 from the acquired image data using the trained machine learning model D2. As described in the first embodiment, there is a correlation between the image data and the feature parameters of the gel structure L2. Furthermore, as shown by the experimental results, according to the second embodiment, the estimation device 100 can accurately calculate the design parameters of the gel structure L2 from the image data.

[0045] The machine learning model D2 according to the second embodiment is configured as a regression model, whereas the machine learning model D2 according to other embodiments may be configured as a classification model.

[0046] The estimation device 100 according to the second embodiment estimates feature parameters from image data, but the estimation device 100 according to other embodiments may obtain both design parameters and feature parameters from image data using both the machine learning model D2 according to the first embodiment and the machine learning model D2 according to the second embodiment. Furthermore, the machine learning model D2 according to other embodiments may be trained to receive image data as input and output each of the design parameters and feature parameters.

[0047] Third Embodiment An estimation device 100 according to a third embodiment outputs the basis of estimation in addition to the estimation result by the estimation unit 105. Fig. 8 is a schematic diagram showing the configuration of the estimation device 100 according to the third embodiment. The estimation device 100 according to the third embodiment further includes a basis identification unit 107 in addition to the configuration of the first embodiment.

[0048] The basis identification unit 107 identifies the basis for the output of the machine learning model D2. The basis identification unit 107 may be realized, for example, by an explainable AI. The explainable AI is realized, for example, by a known library provided as a Python library. For example, the basis identification unit 107 may be implemented by a library such as Grad-CAM or Lime. For example, the basis identification unit 107 may display on an image the magnitude of the gradient of the convolution layer in the calculation by the machine learning model D2. Since a large gradient indicates a large contribution to the estimation result, the basis identification unit 107 can highlight and display pixels that contributed to the estimation result. The output unit 106 outputs the image generated by the basis identification unit 107.

[0049] FIG. 9 is a diagram illustrating an example of a presentation of a basis according to the third embodiment. The example illustrated in FIG. 9 is an image generated by the basis identification unit 107 using Lime. According to the third embodiment, as illustrated in FIG. 9, a region of pixels that have a high contribution rate to the estimation result of the design parameter from the image data is highlighted and displayed as an area surrounded by a white line. This allows the user to confirm which part the trained machine learning model D2 focuses on during estimation. As a result, the estimation device 100 provides an explanation of the basis used during estimation, allowing the user to understand the basis on which the estimation device 100 obtained the estimation result.

[0050] The estimation device 100 according to the third embodiment estimates design parameters and presents the grounds for estimating the design parameters, as in the first embodiment. On the other hand, the estimation device 100 according to other embodiments may estimate feature parameters and present the grounds for estimating the feature parameters, as in the second embodiment.

[0051] Fourth Embodiment FIG. 10 is a schematic diagram showing the configuration of an estimation device 100 according to a fourth embodiment. The estimation device 100 according to the fourth embodiment includes an imaging unit 110 that captures an image of the laminate L. The imaging unit 110 may be a device capable of measuring the laminate L at any scale, such as a laser microscope. The image acquisition unit 101 of the estimation device 100 according to the fourth embodiment controls the imaging unit 110 to capture images under predetermined imaging conditions. Examples of imaging conditions include whether a portion or the entire laminate L is captured, the number of meandering waves captured in the image data, and the contrast of the image. For example, if the image data acquired by the imaging unit 110 does not satisfy the imaging conditions, the image acquisition unit 101 changes parameters of the imaging unit 110 (e.g., magnification, imaging position, etc.).

[0052] The estimation device 100 according to the fourth embodiment can obtain image data captured under certain imaging conditions, thereby reducing the possibility of errors in parameter estimation.

[0053] The estimation device 100 according to the other embodiments may include the basis specification unit 107, as in the third embodiment. The imaging conditions may be determined based on the basis specified by the basis specification unit 107.

[0054] Fifth Embodiment The estimation device 100 according to the first to fourth embodiments estimates parameters of the laminate L from image data. In contrast, the estimation device 100 according to the fifth embodiment estimates the value of another parameter from a combination of image data and parameters. For example, the estimation device 100 estimates the value of a design parameter from a combination of image data and the value of a feature parameter. That is, the explanatory variables of the machine learning model D2 according to the fifth embodiment are the image data and the feature parameter, and the objective variable is the design parameter.

[0055] As described in the first embodiment, there is a correlation between the image data and the design parameters of the gel structure L2, and between the design parameters and the feature parameters of the gel structure L2. Therefore, according to the fifth embodiment, it is possible to estimate the design parameters with higher accuracy than in the first embodiment.

[0056] Note that the estimation device 100 according to other embodiments may estimate the value of a characteristic parameter from a combination of image data and the value of a design parameter. Furthermore, the estimation device 100 according to other embodiments may estimate the value of a second design parameter (e.g., the concentration of a cross-linking agent) from a combination of image data and the value of a first design parameter (e.g., the width of the non-bonded region L12). Furthermore, the estimation device 100 according to other embodiments may estimate the value of a second characteristic parameter (e.g., the hardness of the gel structure L2) from a combination of image data and the value of a first characteristic parameter (e.g., the amplitude of meandering).

[0057] Sixth Embodiment In an estimation device 100 according to a sixth embodiment, the learning unit 104 performs hyperparameter search for the machine learning model D2 and optimizes the machine learning model D2. The hyperparameter search can be realized using, for example, Optuna, a Python library, or the GridSearchCV function of the Scikit-learn library. In the hyperparameter search for the machine learning model D2, the learning unit 104 searches for a learning rate, an optimal function, an intermediate layer, and the like. Note that, since the parameters that can be set in the parameter search vary depending on the structure of the machine learning model D2, the hyperparameters to be searched are determined for each machine learning model D2. According to the sixth embodiment, the estimation device 100 can build a highly accurate machine learning model D2 by searching for hyperparameters during learning.

[0058] Other Embodiments Although one embodiment has been described in detail above with reference to the drawings, the specific configuration is not limited to the above, and various design modifications and the like are possible. That is, in other embodiments, the order of the above-described processes may be changed as appropriate. Furthermore, some processes may be executed in parallel. The estimation device 100 according to the above-described embodiment may be configured by a single computer, or the configuration of the estimation device 100 may be distributed across multiple computers, and the multiple computers may function as the estimation device 100 by cooperating with each other. In particular, the part of the estimation device 100 that performs learning of the machine learning model D2 (e.g., the learning unit 104) and the part that performs estimation using the machine learning model D2 (e.g., the estimation unit 105) may be implemented by separate computers.

[0059] Although the estimation device 100 according to the above-described embodiment uses two-dimensional image data as explanatory variables of the machine learning model D2, this is not limiting. For example, the estimation device 100 according to other embodiments may use a distance image (three-dimensional image data) measured by a laser scanner or the like as explanatory variables of the machine learning model D2. Structural features such as the height and curvature of the non-adhesive portion L22 of the gel structure L2 naturally appear in the distance image.

[0060] The estimation device 100 according to the above-described embodiment is not limited to the laminate L in which a conduit is formed by buckling delamination of a gel structure L2 as shown in Fig. 1. For example, the estimation device 100 according to other embodiments may estimate parameters of a gel structure having a curved portion formed by a method other than buckling delamination.

[0061] <Computer Configuration> FIG. 11 is a schematic block diagram showing the configuration of a computer according to at least one embodiment. The computer 50 includes a processor 51, a main memory 52, a storage 53, and an interface 54. The above-described estimation device 100 is implemented on the computer 50. The operations of the above-described processing units are stored in the storage 53 in the form of a program. The processor 51 reads the program from the storage 53, loads it into the main memory 52, and executes the above-described processing in accordance with the program. The processor 51 also allocates a storage area corresponding to the above-described storage unit 103 in the main memory 52 or the storage 53 in accordance with the program. Examples of the processor 51 include a CPU (Central Processing Unit), a GPU (Graphic Processing Unit), and a microprocessor. The imaging unit 110 according to the fourth embodiment is connected to the processor 51 via the interface 54.

[0062] The program may be for implementing some of the functions to be performed by the computer 50. For example, the program may be implemented in combination with other programs already stored in storage or in combination with other programs implemented in other devices. In another embodiment, the computer 50 may include a custom large-scale integrated circuit (LSI) such as a programmable logic device (PLD) in addition to or instead of the above configuration. Examples of PLDs include programmable array logic (PAL), generic array logic (GAL), complex programmable logic device (CPLD), and field programmable gate array (FPGA). In this case, some or all of the functions implemented by the processor 51 may be implemented by the integrated circuit. Such an integrated circuit is also an example of a processor. In another embodiment, the computer 50 may be virtualized on one or more computers.

[0063] Examples of storage 53 include a magnetic disk, a magneto-optical disk, an optical disk, and a semiconductor memory. Storage 53 may be an internal medium directly connected to the bus of computer 50, or an external medium connected to computer 50 via interface 54 or a communication line. Furthermore, when this program is distributed to computer 50 via a communication line, computer 50 that receives the program may load the program into main memory 52 and execute the above-described processing. In at least one embodiment, storage 53 is a non-transitory tangible storage medium.

[0064] The program may also be a program for realizing part of the above-described functions. Furthermore, the program may be a so-called differential file (differential program) that realizes the above-described functions in combination with another program already stored in storage 53.

[0065] DESCRIPTION OF SYMBOLS 100... Estimation device 101... Image acquisition unit 102... Parameter input unit 103... Memory unit 104... Learning unit 105... Estimation unit 106... Output unit 107... Basis identification unit 110... Imaging unit 50... Computer 51... Processor 52... Main memory 53... Storage 54... Interface D1... Learning dataset D2... Machine learning model L... Laminated body L1... Solid substrate L11... Adhesion region L12... Non-adhesion region L2... Gel structure L21... Adhesion portion L22... Non-adhesion portion

Claims

1. An estimation device comprising: an image acquisition unit that acquires image data showing a gel structure having a curved portion; and an estimation unit that estimates parameters related to the gel structure shown in the acquired image data based on the acquired image data, using a trained model that has been trained to output parameters related to the gel structure shown in the image data from the image data showing the gel structure.

2. The estimation device according to claim 1, wherein the gel structure is provided on a solid substrate and further has an adhesive portion adhered to the solid substrate, and the curved portion is not adhered to the solid substrate and is curved in a direction away from the solid substrate.

3. An estimation method comprising the steps of: acquiring image data showing a gel structure having a curved portion; and estimating parameters related to the gel structure based on the acquired image data using a trained model trained to output parameters related to the gel structure shown in image data showing a gel structure having a curved portion.

4. A method for generating a trained model, comprising: a step of acquiring a training dataset including a combination of sample image data depicting a gel structure having a curved portion and parameters related to the gel structure depicted in the sample image data; a step of learning internal parameters of a machine learning model using the training dataset so as to output parameters related to the gel structure from image data depicting the gel structure; and a step of generating a trained model using the trained internal parameters, wherein the trained model is generated by causing a computer to execute the steps of: accepting input of image data; and calculating parameters related to the gel structure depicted in the image data based on the input image data and the internal parameters.

Images